An efficient depth estimation using temporal 3D-Warping

This paper presents a computationally efficient method for estimation of high-quality depth for multiview video acquired by a camera array in motion. Depth information is essential for 3DTV display system for generating video streams from virtual view points. Dense depth estimation has been successfully modeled as a Markov Random Field, and several methods like Iterated Conditional Modes, Graph Cuts and Belief Propagation have been proposed to solve the same. While depth estimation using Graph Cuts and Belief Propagation give accurate results, their computational requirements are high. On the other hand, Iterated Conditional Modes is fast, but the quality of the result is poor. In our work, we propose a technique for boosting the quality of the resultant depth estimated using Iterated Conditional Modes to near Graph Cuts or Belief Propagation levels while keeping the computational cost low.